Assessment of naturalness in the prosynth speech synthesis project

Abstract

This paper describes the approach to the assessment of the naturalness of synthetic speech taken within the ProSynth collaborative speech synthesis project. The view expressed is that an important aspect of naturalness is ease of understanding, and the consequences are that this leads to a means for the evaluation of scientific hypotheses through perceptual tests. The premise within ProSynth is that listeners' processing of synthetic speech will be faster and more accurate when the signal includes phonetic fine detail that systematically varies with the linguistic structure. Four perceptual experiments are outlined which demonstrate both the application of the approach and the effectiveness of the basic principle.